Skip to content

QECTOR v1.0.6

Latest

Choose a tag to compare

@github-actions github-actions released this 24 Aug 00:12
· 16 commits to main since this release

QECTOR Claude Plugin v1.0.6

Release date: 2026-08-23
Release type: Claude.ai marketplace compliance hardening + setup robustness
Full changelog: v1.0.5...v1.0.6


Why this release exists

v1.0.5 introduced the cross-platform launcher in a bin/ directory. When the
claude.ai-hosted marketplace pipeline (the one Claude Desktop and Claude Code
sync from on hosted/enterprise setups) reviewed the plugin, it flagged a
hard compliance problem: claude.ai-hosted plugins may not ship bin/
executables
. Because bin/ entries are added to the command PATH when the
plugin loads, but are never shown on the admin approval surface, a plugin
shipping one represents a silent code-execution surface that administrators
cannot see or approve. This is an Anthropic platform policy, not a defect in
the launcher itself.

One more correctness gap surfaced during the same review window: the
/qec-setup command hard-required a local scripts/qector_system_setup.py.
On a normal machine that works. But in the sandboxed and remote execution
environments Claude also offers (claude.ai code execution, Cowork, containers
without a project checkout), that script does not exist, so the command
dead-ended. That is now fixed, not just patched.


Scope and impact

Who this release affects:

Audience Impact
Everyone Launchers relocated from bin/ to scripts/ in every artifact type. If you installed v1.0.5, reinstall the v1.0.6 bundle — do not reuse v1.0.5 cache files.
claude.ai-hosted / marketplace users This was the direct trigger. The manifests are now canonical and the plugin loads in the strictest approval surfaces.
Windows users No functional change. The Windows launcher is selected via the win32 override exactly as in v1.0.5 — it just lives in scripts/ now.
Sandboxed / remote users /qec-setup no longer dead-ends; it falls back to native diagnostics and reports honestly.

What changed in v1.0.6

1. Claude.ai marketplace compliance (fixes a platform-policy rejection)

  • Launchers moved from bin/ to scripts/. Both qector-python (POSIX
    sh) and qector-python.cmd (Windows) now live in scripts/, with the exec
    bit preserved through packaging (git mode 100755, archive attr 0o755).
    Every reference was repointed: plugin.json, .mcp.json, hooks/hooks.json
    (SessionStart + PostToolUse), the Desktop MCPB manifest, the win32
    platform_overrides block, the builder whitelists, and the bundle validator.
    Commit 2bd1aa4.
  • Canonical marketplace manifests. marketplace.json now declares the
    plugin with the relative same-repo source ("source": "./"), the form every
    official Anthropic marketplace uses and the only form guaranteed resolvable
    across every sync path that clones the repository. All fields are from the
    documented schema (homepage, repository, keywords, owner url).
    Commits 5814242 and 7bb49f7.
  • bin/ ban guard (regression lock). validate_plugin_bundle.py now
    hard-fails if a bin/ entry appears in any of the three release artifacts
    (plugin zip, source zip, Desktop MCPB). Commit 2bd1aa4.

2. Environment-agnostic setup

  • /qec-setup works everywhere. The command now detects whether it is
    running on a real checkout (script present) or in a sandbox/remote/no-clone
    environment, and falls back to native diagnostics (Python version, pip,
    package versions, artifacts directory) instead of failing. It explicitly
    instructs Claude to never claim the script ran when it used the
    fallback — honest reporting by design. Commit c49b34f.

Interpreter pinning note (no action unless you set it in v1.0.5)

The userConfig.python_path / user_config.python_path block was removed
from plugin.json in this release. This was the newest and least-standard
manifest field, and the strict claude.ai approval schema does not recognize
it. The feature is not lost: interpreter pinning still works end to end —
set the QECTOR_PYTHON environment variable to an absolute Python 3.9-3.13
path and every launcher honors it before auto-resolution. If you previously
configured userConfig.python_path, migrate by setting QECTOR_PYTHON
instead; there is a one-line note in the README.

Supported environments

  • Python: 3.9 through 3.13. Launchers explicitly reject 3.14+ (the
    qector-decoder-v3 wheel matrix does not cover it) with guidance, rather
    than crashing at import time.
  • qector-decoder-v3==1.0.0, the pinned decoder runtime.
  • Claude Code: Windows, macOS, Linux (via the shipped launchers).
  • Claude Desktop: Windows, macOS (Desktop MCPB, safe profile).
  • Web / iOS / Android / Cowork: local stdio cannot reach those surfaces
    (Anthropic architecture); a hosted remote connector is on the 1.1.x
    roadmap.

What does NOT change

  • Surface counts. 28 skills, 11 commands, 5 agents, 4 MCP servers, 8
    stable library tools, 32 research tools, 3 admin tools — identical to v1.0.5.
  • All 16 reference theorems remain executable-proof-verified in the test
    suite (74 unit tests, 48 subtests across the math, protocol, and production
    readiness gates).
  • The eight-tool safe Desktop profile is unchanged.
  • Offline-by-default privacy binary is unchanged (no network calls unless
    explicitly requested).

Known limitations

  • Sandbox / remote diagnostics depth. The native fallback path in
    /qec-setup reports the core audit (interpreter, pip, package versions,
    artifacts dir) but does not run the full math validation the local script
    performs (that needs the qector-decoder-v3 wheel, which sandboxes may not
    have). It is intentionally reduced scope, clearly labeled, and sessions are
    told to install and validate on a real machine when possible.
  • Marketplace sync still needs a private repo on managed orgs. Per the
    Claude documents, if you use Claude Desktop's managed-org marketplace
    pipeline (the one that says "Sync automatically"), the repository must be
    private or internal and owned by the same org as the Claude GitHub App. A
    public repo can still be rejected with a generic sync error on that
    pipeline, even though every direct install path works.

Integrity

Artifacts are built deterministically with fixed timestamps, so the build is
byte-for-byte reproducible. SHA-256 hashes are published in SHA256SUMS,
sidecar files, and the MCP Registry descriptor, and the registry entry's
fileSha256 matches the published Desktop MCPB.

Artifact SHA-256
qector-claude-desktop-1.0.6.mcpb 5b6b4c247ef6159dc92441023fd08a0dc800894a220ba8a61b4143bde92190ff
qector-claude-plugin-1.0.6.zip e1660a45e87e62d5b74f561273ff7cac2a5367a70b8418f1f9c80ea69591de7f
qector-claude-plugin-source-1.0.6.zip e0e04d94546799acda0d8f3258bc8911b8125c4bf2b78bb80c04d3e49588dc59

Verification summary: 832/832 source checks · bundle ALL CLEAR · 15/15
release metadata · 74/74 unit tests (+48 subtests) · claude plugin validate . --strict passes · fresh marketplace add + install exits 0.


Upgrade

Claude Code: /plugin marketplace update qector-tools then
/plugin install qector@qector-tools.

Claude Desktop: remove the old v1.0.5 extension, download
qector-claude-desktop-1.0.6.mcpb from the release page, install via the
extension UI, restart when prompted.


Traceability

Commit Change
5814242 Canonical marketplace manifests
09ca506 Drop unknown license fields (strict sync)
2bd1aa4 Launchers bin/scripts/, bin/ ban guard
c49b34f /qec-setup environment-agnostic fallback
af8a2ac v1.0.6 release commit

Acknowledgments

Thanks to the claude.ai marketplace review pipeline for surfacing the bin/
approval-surface policy, which drove the relocation, and to everyone who
exercised the marketplace and /qec-setup paths from real Desktop and
sandboxed environments and reported the gaps. Straightforward, reproducible
feedback like that is what makes the resilience layer worth keeping.


Questions, edge-feature requests, or QEC research collaborations: open an
issue or reach out via the repository. QECTOR is source-available; the
reference manual ships separately via Zenodo DOI.